Skip to content

zentorch Release v2.12.0.0

Latest
Latest

Choose a tag to compare

View all tags
@kiriti-pendyala kiriti-pendyala released this 20 May 15:35
· 45 commits to main since this release
v2.12.0.0
f2d1555
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Versioning Change

Starting this release, zentorch adopts a PyTorch-aligned versioning scheme: MAJOR.MINOR.PATCH.PLUGIN_PATCH. The first three segments match the PyTorch version; only PLUGIN_PATCH is maintained by zentorch. Weekly development builds are published separately as zentorch-weekly on PyPI.

Previous This Release
zentorch 5.2.1 2.12.0.0

Compatibility Matrix

Dependency Supported Versions
PyTorch 2.12.0, 2.11.0
ZenDNN 5.2.2
Python 3.10 – 3.13
vLLM 0.20.2, 0.20.1, 0.20.0
TorchAO 0.17.0, 0.16.0

Deprecations & Removed Support

  • PyTorch 2.10.0 support dropped — minimum required version is now 2.11.0 (N-1 policy).
  • vLLM < 0.20.0 no longer supported — versions 0.15.0 through 0.19.1 have been removed.
  • Dual-install guard added — importing zentorch now raises ImportError if both zentorch and zentorch-weekly are installed simultaneously.

New Features

  • Fused MoE (Mixture-of-Experts) operator — new zentorch op for MoE inference.
  • Group MatMul integration — core group matmul support in zentorch.
  • ZenDNN SDPA (Scaled Dot-Product Attention) — integrated as a zentorch op.
  • QKV fusion pass in export path — enables fused QKV projections during torch.export.
  • Linear AOT Inductor (AOTI) integration — linear ops now work with torch._export + AOTI.

Bug Fixes

  • Added prepack and post-op validation checks in QKV fusion pass.
  • Added feature gate to disable fp16 op replacement when not applicable.
  • Fixed PyTorch 2.12 inductor hook signature and test dependencies.
Assets 2
Loading